Hebrew Vowel Restoration With Neural Networks

Modern Hebrew is written without vowels, presenting a problem for those wishing to carry out lexical analysis on Hebrew texts. Although fluent speakers can easily replace vowels when reading or speaking from a text, there are no simple rules that would allow for this task to be easily automated. Previous work in this field has involved using statistical methods to try to solve this problem. Instead we use neural networks, in which letter and morphology information are fed into a network as input and the output is the proposed vowel placement. Using a publicly available Hebrew corpus containing vowel and morphological tagging, we were able to restore 85% of the correct vowels to our test set. We achieved an 87% success rate for restoring the correct phonetic value for each letter. While our results do not compare favorably to previous results, we believe that, with further experimentation, our connectionist approach could be made viable.